Qwen3.8-27B vs Muse Glimmer 30B: Best New Model for a 24GB Card?

Short answer

Choose Qwen3.8-27B if your priority is:

• more VRAM headroom on a 24GB card (~16 GiB weights at Q4)
• long context: 262,144 native, extensible to 1M via YaRN
• software-engineering agents (SWE-bench Pro, Terminal-Bench)
• day-0 GGUF availability

Choose Muse Glimmer 30B if your priority is:

• planning and tool-use quality in agentic loops
• fast drafting via the DFlash speculative decoder
• on-device deployment (ExecuTorch builds ship in the collection)
• a dense model with predictable per-token behavior

Both are Apache 2.0. Both are multimodal. The 24GB card runs either one.

Four days apart, same target

Qwen announced Qwen3.8-27B on August 3, 2026, and the weights landed on Hugging Face August 13-14. Meta announced Muse Glimmer 30B on August 10, 2026. That is two Apache 2.0 multimodal models, released four days apart, aimed at the same single-GPU buyer.

The Meta release is the bigger story on license alone. VentureBeat headlined it as “Meta returns to open source.” Glimmer is the first Apache-licensed Meta model in this class — no custom community license, no acceptable-use gate.

Specs that matter

VRAM: both fit 24GB, one fits easier

Qwen3.8-27B at Q4_K_M is 15.93 GiB of weights. Plan on roughly 18-20 GiB total with 32-64K context. On a 24GB card, that leaves room for a long agent session. FP8 needs about 45 GiB with native-context KV cache, and BF16 needs about 68 GiB — those tiers belong on the 64GB and 128GB unified-memory lists.

Glimmer’s language model is under 20GB at 4-bit. Add the perception encoder and KV cache and Meta targets the 24-32GB envelope. It fits a 24GB card, but with less slack.

Unsloth published day-0 GGUFs for Qwen3.8-27B. That is convenient, but quant quality is unvalidated in the first week — wait for community perplexity checks before you trust a specific file for real work.

Benchmarks: two vendors, two stories

All launch numbers below are vendor-reported. No independent reproduction existed in the first week.

Qwen’s own comparison table shows Qwen3.8-27B leading every overlapping text row against Glimmer — but the vendor picks which rows overlap. Against its own predecessor, Qwen reports large jumps: Terminal-Bench 2.1 goes 63.4 → 73.0, DeepSWE 1.1 goes 13.3 → 42.2, SWE-bench Pro goes 53.5 → 61.7, and OSWorld-Verified goes 63.9 → 84.3.

Meta’s numbers point the other way on agentic work. Coverage of the launch reports Glimmer outperforming Gemma4-31B and Qwen3.6-27B on agentic benchmarks, while trailing on computer-use and terminal tasks. So the split is roughly: Qwen3.8 for software-engineering agents, Glimmer for planning and tool-use loops. Treat both claims as marketing until independent numbers land.

For how the previous generation compared, see Qwen3.6 vs Gemma 4 for agentic work.

The DFlash factor

Glimmer ships with DFlash, a block-diffusion drafter that proposes 16 tokens per forward pass for speculative decoding. Meta also ships ExecuTorch builds in the Hugging Face collection. Together they signal what Glimmer is for: on-device agents where decode speed and deployment footprint matter more than leaderboard rows.

Qwen3.8’s answer is architectural: 48 of its 64 layers use Gated DeltaNet linear attention, which keeps long-context prefill affordable. That is what makes the 262K native context usable rather than theoretical.

Decision table

Final recommendation

For most OpenClaw users on a 24GB card: start with Qwen3.8-27B. It is smaller at Q4, its context story is stronger, and the GGUF path is already live.

Test Glimmer 30B if your agent spends more time planning and calling tools than editing code, or if you want the DFlash decode path. And watch the independent benchmarks — every number above came from a vendor launch post.

Next steps

• Best local models for OpenClaw
• MoE vs dense on a 24GB card
• Best local LLMs for 32GB RAM

Sources

• Qwen3.8-27B model card (Hugging Face)
• Muse Glimmer 30B model card (Hugging Face)
• Unsloth Qwen3.8-27B GGUF